Pre-match trading: gather external context for scheduled games instead of polling Shipp

[buildkit-ai]

Problem

The current TradingLoop.run() immediately enters a live event polling loop regardless of game status. For scheduled (not yet started) games, this means:

1. Shipp's getLiveEvents() returns empty — the game hasn't started, there are no events
2. The loop just sleeps and retries, doing nothing until tipoff
3. Pre-game markets are actively traded on Kalshi but the bot ignores them entirely

This is a missed opportunity. Pre-game markets often have the most liquidity, and edges can exist before games start.

Proposed Solution

Split the trading loop into two distinct phases: pre-match and live.

Phase 1: Pre-match (game status = scheduled)

Instead of polling Shipp for nonexistent live events, gather context from other sources:

Data sources for pre-match context:

• Shipp metadata (already available): team rosters from home_team_players / away_team_players stored in games.metadata
• News headlines: recent team news via RSS or news APIs (injuries, trades, lineup changes)
• Odds/lines: opening lines from sportsbooks as a baseline reference
• Team stats: season record, recent form (last 5/10 games), home/away splits
• Head-to-head history: recent matchup results between the two teams
• Injury reports: official injury reports (out/doubtful/questionable)

Architecture:

interface PrematchDataSource {
  name: string
  gather(game: Game, sport: string): Promise<string>
}

A pluggable system where data sources are registered and called in parallel during pre-match. Each returns a text block that gets concatenated into the full pre-match context. This makes it easy to add new sources without modifying the core loop.

Built-in sources to start:

• ShippMetadataSource — extracts rosters/venue from stored game metadata
• NewsHeadlineSource — fetches recent headlines via Google News RSS
• KalshiMarketSource — includes current market prices as implied probabilities

Future sources (community contributions):

• ESPNStatsSource, NBAStatsSource — team/player statistics
• OddsAPISource — consensus lines from sportsbooks
• InjuryReportSource — official injury designations
• WeatherSource — for outdoor sports (NFL, MLB, Soccer)

Phase 2: Live (game status = live)

Once the game starts, switch to the existing Shipp live event polling loop. No changes needed here.

AI Adapter Changes

Add a estimatePrematch() method alongside the existing estimateProbability():

• Different system prompt tuned for pre-match analysis (base rates, team strength, home/away advantage)
• Accepts a free-text prematchContext string instead of ShippEvent[]
• Same structured output: { yesProbability, confidence, reasoning }

Risk Management Considerations

Pre-match trades may warrant different risk parameters:

• Higher minimum edge threshold (markets are more efficient pre-game)
• Lower confidence acceptance (less information available)
• Optional strategy field to distinguish pre-match vs live orders in the DB (e.g., value-bet-prematch vs value-bet)

Flow

Game status = scheduled
  ├── Gather pre-match context (rosters, news, stats, market prices)
  ├── For each Kalshi market:
  │   ├── AI estimates probability from pre-match context
  │   ├── Compute edge vs market price
  │   ├── Risk check → place trade if edge exists
  │   └── Log analysis + reasoning
  ├── Log stats summary
  └── Poll for game status change (scheduled → live)
       └── When live → switch to existing live event loop

Game Lifecycle

scheduled ──[pre-match analysis]──→ waiting ──[game starts]──→ live ──[live loop]──→ completed

Context

Currently the only data flowing into the AI is Shipp's live event stream. By adding pre-match context gathering, the bot can:

• Trade earlier (hours before tipoff when markets open)
• Use a broader information set than just in-game events
• Potentially find larger edges pre-game when markets haven't fully adjusted to news

Related

• Event matching fails: Kalshi uses 'at' separator but code splits on 'vs' #3 — Event matching between Shipp and Kalshi (team name format mismatch)